docs/METRIC_VALIDITY.md

# Metric validity

Metrics are evaluated per scenario and category. Current-belief and authority
metrics use five scenario rows per category. Historical metrics use only rows
with an expected historical answer. Promotion, erasure, temporal, and
evidence-family metrics return `not_applicable` when their applicability
denominator is zero; they do not award a perfect score for an empty strategy.
Rates are never inferred from prose. Privacy leakage counts a violation only
when a record is promoted outside its permitted scope. Abstention is evaluated
against an explicitly expected ambiguity decision.